Er, scratch that. I guess the board doesn't have enough power when there's a
second GPU in another PCIe slot. It comes up fine now, and I get the same
issue.

Looks like this bit of nvkm_ioctl_new is somehow failing with -ENODEV. My
latest theory is:

nvkm_fifo_chan_child_new calls engine_ctor (nv40_fifo_dma_engine_ctor), which
in turn calls nvkm_object_bind() on something it's not supposed to (like the
engine object, I think), which in turn returns -ENODEV as there's no bind
pointer. I suspect the solution here is to add a dummy .bind to nv31_mpeg_chan,
since the binding effectively happens at chan_new time. Or we could move the
mpeg->chan check to the bind action.